In November 2025, three Stanford economists pulled monthly payroll records from ADP and found something the broad “AI is coming for everyone” story doesn’t quite capture. Employment for 22-to-25-year-olds in the most AI-exposed occupations fell roughly 13% after controls. Workers aged 35 to 49 in those same jobs were barely touched, and in raw terms actually grew.
Same roles, same tools, opposite outcome by age. The early strain isn’t spread evenly. It’s landing on the young.
A quick note before I go further. I’m a writer with an interest in this, not an economist, and what follows is my reading of a single working paper, not career advice. The study here is early, and it shows a link, not proof. It’s a finding from one dataset, not a settled verdict about anyone’s future.
What the payroll data actually found
The three Stanford economists behind the paper, Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, gave it the slightly ominous title “Canaries in the Coal Mine?”. Instead of surveys or forecasts, they used something harder to argue with: monthly payroll records from ADP, the largest US payroll processor, with the company’s services covering over 25 million workers. Their main analysis sample tracks 3.5 to 5 million of those workers each month from January 2021 through September 2025.
The headline finding sits at the very top of the abstract: workers aged 22 to 25 in the most AI-exposed occupations, software development, customer service, clerical work, saw a 16% relative decline in employment, after controlling for firm-level shocks. Look at the raw numbers without those controls and the same group is down about 6% from late 2022, while workers aged 35 to 49 in the same exposed jobs are up over 8% across the same period. For software developers specifically, the picture is sharper: 22-to-25-year-olds in that occupation are down nearly 20% from their late-2022 peak.
Same jobs, same tools, opposite outcome by age. Companies weren’t cutting salaries. The paper finds little difference in annual salary trends across age or exposure groups. They were hiring fewer juniors.
Why the age split makes a grim kind of sense
The explanation the authors lean on inverts an assumption a lot of us carry. An AI model learns from the written record: books, articles, documentation, all the material a graduate spends years absorbing. As Brynjolfsson put it, “That’s the kind of book learning that a lot of people get at universities before they enter the job market, so there is a lot of overlap with between these LLMs and the knowledge young people have.”
Some might think experience gets automated first, the boring repeatable stuff, and that fresh credentials are the safe currency. The data suggests the opposite. The textbook knowledge a junior is hired on is exactly what the model already has.
What the model doesn’t have is the stuff nobody wrote down. Brynjolfsson frames it this way: “Older workers have a lot of tacit knowledge because they learn tricks of trade from experience that may never be written down anywhere. They have knowledge that’s not in the LLMs, so they’re not being replaced as much by them.” That’s his read of the pattern, not a proven cause, and I’d hold it loosely, but it fits more cleanly than most explanations of the data manage.
The paper layers in a useful nuance here. Not every use of AI is associated with declines: when the authors split occupations by whether AI is being used to automate tasks versus augment them, the entry-level drops are concentrated in the automation side. Where AI is mostly augmenting human work, young workers’ employment has held up or grown. It seems the technology isn’t a single force flattening everything. Its effect depends on how it’s actually being used at work.
It also isn’t only this one paper. Harvard economists Seyed Hosseini Maasoum and Guy Lichtinger, using résumé and job-posting data covering roughly 62 million US workers across 285,000 firms, found that at firms adopting generative AI, junior employment declined sharply relative to non-adopters starting in 2023, while senior employment in the same firms kept rising.
The millennial caught in the middle
I read this as a millennial, which is a strange seat to read it from. We’re not the canaries here. The 22-to-25 bracket is the cohort behind us. But I don’t think anyone my age should feel safe by a few years’ margin.
The career paths a lot of us committed to are going obsolete faster than the timeline we signed up for. We’re old enough to have picked a direction and put a decade into it, and young enough to have another two or three decades of working life left, with no real option but to keep adapting the whole way. That’s just the shape of the thing. A skill set seems to go stale faster than it used to, and you don’t get to opt out of that by having a head start.
If there’s a useful line in the paper for anyone in this position, it’s the one where Brynjolfsson stops describing and starts advising. “Young workers who learn how to use AI effectively can be much more productive,” he said. “But if you are just doing things that AI can already do for you, you won’t have as much value-add.” That’s opinion sitting next to data, and it sounds right to me.
What this changes about thinking on a career
I want to be careful not to turn one study into a doctrine. Brynjolfsson himself reaches for the long view, noting that “Tech has always been destroying jobs and creating jobs. There has always been this turnover.” History is full of automation panics that didn’t end the way the panic predicted. Whether AI follows that script or breaks it is the open question, and the authors themselves caution that the patterns they observe may in part be influenced by factors other than generative AI.
What I take from the data is smaller and more personal. A foothold early in a career is shakier than we assumed, and the order in which we thought things were safe may be backwards. The entry rung, the one you climb onto with a credential, is where the pressure is showing up first.
I have some lived sympathy for that wobble. Around 30, I left running an adult language school in Vietnam to take a VC internship, while friends my age were already qualified accountants comfortably on a track. The status drop was real, and it also worked out fine. Starting from a weaker rung isn’t the same as being finished, just slower and less comfortable than the version where the ladder stays put.
The honest answer is that we don’t yet know which story this is. One reading is that juniors are simply the first to feel a pressure that will work its way up the ladder as the models get better at tacit, on-the-job knowledge too. The other reading is that this is an adjustment, painful at the entry rung but localised there, and the rest of the workforce reorganises around it without the same shock. The payroll data is consistent with both. It’s going to take a few more years of numbers, and probably a few more papers like this one, before anyone can say which way the canary was pointing.
